In order for brain-inspired computing to become a reality, the underlying hardware must become sufficiently powerful to do in-silicon what the brain does naturally. One of the important advances in this field was the 2014 debut of the first neuromorphic chip, “True North,” developed by IBM with funding from the Defense Advanced Research Projects Agency.
This week, a project was announced that will further explore the technology that was set in motion by IBM. As part of the “Massively Parallel Modeling and Simulation of Next-Generation Hybrid Neuromorphic Supercomputer Systems,” researchers at the Rensselaer Polytechnic Institute will address the potential for the neuromorphic processor to incorporated into a next-generation supercomputer.
With a $1.3 million grant from the Air Force Research Laboratory, the Rensselaer team will use the institute’s supercomputer, Amos, to run a massively parallel simulation of proposed neuromorphic supercomputer designs. The model will also enable them to examine various network designs and assess the machine’s suitability for compute- and data-intensive science and engineering problems.
The researchers are looking to test the feasibility of building a hybrid supercomputer that uses both neuromorphic and conventional processors. The mixed-processor approach been growing in popularity ever since Roadrunner broke the petaflops barrier in 2008 with AMD x86 CPUs and souped-up IBM Cell coprocessors. Although this is a simulation and will not use the actual True North hardware, the researchers will base the model on IBM’s specifications for the processor and its simulation development kit.
“The question we’re asking is: What if future supercomputer designs were to have several embedded neuromorphic processors?” said Christopher Carothers, director of the Center for Computational Innovations, in the official announcement. “How would you design that computer? And what new capabilities would it offer?”
One of the primary advantages of the brain’s computational ability is its speed and efficiency. Computers that rely on neuromorphic processing are expected to use far less energy and emit less heat than conventional chips. They will also excel at classification problems, like pattern recognition and dealing with error messages.
The True North architecture supports very fast parallel computing that is naturally fault-tolerant. The chip incorporates 5.4 billion transistors arranged in a network of 4,096 neurosynaptic cores, yielding the equivalent of one million neurons and 256 million synapses.
Depending on how this and other research pans out, neuromorophic processors could one day soon provide another tool in the heterogenous computing toolbox, alongside general-purpose GPUs and Intel’s Many Integrated Core (MIC) architecture.